DEV Community

Cover image for What Makes a Cloud a Cloud? The Five Stamps That Decide
Amar
Amar

Posted on Originally published at iqraa.tech

What Makes a Cloud a Cloud? The Five Stamps That Decide

TL;DR: What makes a cloud a cloud is a test in one NIST sentence from September 2011: computing capabilities drawn from a shared pool, reached over a network on demand, released with minimal management effort, and metered, and a service earns the word only when all five stamps that follow from it hold.

  • NIST SP 800-145 (September 2011, authors Peter Mell and Tim Grance) holds the reference definition, and its load-bearing phrases are “shared pool” and “minimal management effort”; a paraphrase that drops either is describing rented servers.
  • The five stamps: on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service. A candidate service passes only with all five.
  • Security, automation, resilience, budget-friendliness, and flexibility are badges: outcomes the stamps produce, never parts of the definition.
  • Scalability is designed headroom; elasticity adds automatic shrinking, so a fixed monthly server can scale and still fail stamp four.
  • Amazon S3’s SLA pays a 10 percent credit below 99.9 percent monthly uptime, and egress past the first free 100 GB a month runs about $0.09 per GB.

The label “cloud” costs a vendor nothing to print, so the lesson embedded above replaces it with a test that costs something to pass. This article is the companion to that Whiteboard Architect episode: watch the full episode on YouTube to see the boards drawn, or read on: same claims, sources named, plus the tables and fine print a whiteboard cannot hold.

Whiteboard cold open titled Everything wants to be cloud: a hand drawn cloud tag crossed out in red beside a sketch of three stacked servers, a blue badge reading 5 stamps, and a red bordered note reading all five, pass or fail

The cold-open board: the label anyone can print, and the five stamp test that decides.

The sentence that decides, and its two load-bearing phrases

Peter Mell and Tim Grance at NIST published Special Publication 800-145 in September 2011, three pages that still arbitrate the word. The model it describes: capabilities over a network whenever wanted, from a pool shared among tenants, handed back with barely any management effort, metered throughout.

Two phrases in that sentence do the discriminating, and both are checkable. A “shared pool” means no machine carries your name plate: the provider reassigns resources as demand moves, and the strongest location promise is a country or a datacenter, never a rack. “Minimal management effort” means provisioning happens without a human on the provider’s side: you click a console or call an API, and the capacity exists minutes later.

Paraphrase carelessly and the test disappears. A common summary compresses it into delivering IT resources over the Internet: the convenience survives, both phrases go missing, and a rented dedicated server starts calling itself a cloud. The sentence also travels: a later ISO standard carries a compatible definition, and contracts and textbooks worldwide follow it.

What does the NIST sentence say, word for word?

The spine of SP 800-145’s definition, quoted in fragments: computing capabilities are available as “a shared pool of configurable computing resources” that can be “rapidly provisioned and released” with “minimal management effort or service provider interaction.” The full sentence opens with “ubiquitous, convenient, on-demand network access” to that pool, adds the parenthetical resource list, “e.g., networks, servers, storage, applications, and services,” and sits on page 2 of the paper’s three-page body.

The clause paraphrases drop first is the interaction one. It limits the humans on the provider’s side of each request, so a portal whose every action waits on a sales engineer fails the sentence before a single stamp is checked. Quoting the paper is also the cheapest defense against drift: both phrases live in one sentence a reader can hold against any brochure.

How many deployment models does the standard count?

Four: private, community, public, and hybrid. Summaries that stop at three have usually dropped community, the pool shared by several organizations with requirements in common, for example agencies under one regulatory regime. Hybrid keeps dedicated and rented capacity joined, so a privacy-bound workload stays on the dedicated side while a seasonal surge rents.

Model Who the pool serves
Private One organization, on infrastructure set aside for it
Community Several organizations with requirements in common
Public Any tenant, on published terms
Hybrid A mixture of dedicated and rented, operated as one

The five stamps that decide what makes a cloud a cloud

SP 800-145 lists exactly five essential characteristics; the episode treats them as stamps, and the count has to reach five.

Stamp The demand it makes A service shape that fails it
On-demand self-service You provision compute, storage, or network capacity yourself, by console or API An order form that emails a human account team
Broad network access Capabilities travel by ordinary protocols to a phone as readily as a workstation A cluster reachable only from machines inside one office
Resource pooling Physical and virtual capacity in one multi-tenant pool, rebalanced as demand moves, with location promises stopping at datacenter or country A host that pins you to one named physical box
Rapid elasticity Capacity expands with a spike and contracts after it, sometimes with no operator involved Fixed monthly capacity, even rented and online in minutes
Measured service Usage monitored, controlled, and reported in units that fit the service, visible to both sides A flat monthly price with no meter behind it

Stamp four is the one advertisements lean on hardest. A team that doubles its instance count for a launch and leaves the extras running afterward has scaled; it turns elastic only when the extra instances terminate themselves once traffic falls. Rented capacity faces the same test, and the discriminating half is the shrink.

Stamp five makes the arrangement a utility: infrastructure bills by allocation and consumption, software by subscription or by use, and a capital budget becomes an operating line that tracks usage. Compute meters in instance-hours and storage in GB-months; both parties read the same numbers, which is what makes an invoice arguable line by line.

What the stamps decide, and what the brochure sells

Everything marketed around cloud computing beyond the five stamps is an outcome. The board splits the vocabulary into two boxes: stamps settle whether a service is a cloud at all, badges score how good a cloud it is. Security, automation, resilience, budget-friendliness, and flexibility sit in the second box; none appears in the paper’s characteristics.

Each badge traces to a stamp or to engineering on top of the definition.

Brochure badge Where it actually comes from
Security Configuration plus shared responsibility; the definition is silent on it
Automation The shape minimal management effort takes at fleet scale
Resilience Redundancy engineering, passive or active, on pooled capacity
Budget-friendliness The meter, plus a workload shaped to suit it
Flexibility Elasticity plus self-service: change allocation yourself, any hour

Whiteboard board 6 of 15 titled Stamps, not badges: on the left a box labeled stamps holding five numbered circles with a green note reading decides: is it a cloud; on the right a box labeled badges, the brochure, listing security, automation, resilience, budget-friendliness, and flexibility under a red note reading none is the test

Board 6 of 15: the stamps settle what the thing is; the badges score how good it is.

Where the two usual impostors fail

Two impostors wear the word most often. An on-premises data center is hardware bought and staffed for the busiest hour of the year; the meter fails first, since no per-use bill exists, and self-service usually goes with it, since provisioning runs through a ticket queue. A dedicated server rented online in minutes sits closer: network reachable, quick to order, but fixed. Nothing shrinks the allocation once traffic drops, and the monthly fee is flat.

The claim Stamps it fails Why
A “cloud” private data center 1 and 5, commonly No self-service path, no per-use meter; capacity is bought for the peak
A “cloud” rented dedicated server 4 and 5 Fixed allocation, flat invoice; capacity never follows load
A cloud by the definition None Self-service, network reach, pooled, elastic, metered: five for five

The logo’s sincerity never enters the test: excellent hardware can still stop one stamp short.

Can a private deployment pass the stamps?

Yes, and the definition never asks who owns the plant. The five characteristics and the four deployment models are separate axes in SP 800-145, so a private cloud built on an organization’s own virtualized hardware qualifies whenever it offers the same pattern: a self-service portal, a pooled substrate, allocation that follows demand, and an internal meter. The checklist from this episode applies to it with no edits, which is also why “we bought our own machines” settles nothing by itself.

Two words the brochure merges: scalable and elastic

Scalability is a system’s ability to absorb a growing workload when resources are added; it is headroom an architect designs in, and the adding can stay manual. Elasticity is stricter: the platform provisions and releases capacity on its own, tracking demand closely in both directions, and the difference matters once a bill is attached. A rented server scales, since a second one is one order away; hours after the spike has passed, the invoice still lists two servers. Elasticity removes the second line.

A later episode gives the pair its own treatment. The question to put to any claimed cloud: when demand falls, do the capacity or the bill fall with it, absent a human in the loop?

Automation, resilience, and the energy bill

Automation never appears as a sixth stamp: it is a consequence. Minimal management effort, applied across a provider’s fleet, turns into software doing the installing, configuring, and maintaining. Orchestration, the wider layer, aligns entire workflows, joining configuration, coordination, and deployment in service of larger goals; automation on its own runs one repetitive task, such as provisioning a volume or applying an update. The board draws the relationship as nested boxes: the single task inside, the whole workflow around it.

Whiteboard board 9 of 15 titled Automation and orchestration: a small box reading automation, one repetitive task, provisioning, updates, nested inside a larger box reading orchestration, whole workflows, with a green arrow pointing toward larger goals and side notes reading minimal manual labor and a consequence, not a sixth stamp

Board 9 of 15: one repetitive task in the inner box, the workflow around it in the outer.

Where exactly is the line between automation and orchestration?

The size of the job. Automation removes people from one repetitive task: creating an account, provisioning a volume, patching an operating system. Orchestration runs the sequence: it invokes those automations in order, carries state between the steps, and unwinds the work when a step fails. The board’s inner box holds the first kind of job, the outer box the second, and a green arrow connects the small one to the phrase larger goals. Neither box is a stamp; both are what the fifth and first stamps look like from inside the provider’s building.

Resilience, also a badge, reduces to redundancy engineering. Providers keep copies of data and of components across many servers and sites, so the loss of one machine, or of a whole site, removes capacity while the service keeps running. Passive redundancy is capacity held before anything breaks, the way a cluster at half load can lose a node and keep answering, slower; active redundancy is machinery that notices a failure and reroutes the work on its own.

Where do the cloud energy numbers come from?

An IEA estimate put global data center electricity use at roughly 1.5 percent in 2024. Hyperscale operators raise utilization through consolidation and virtualization, site buildings near renewable supply, cool with outside air, reuse waste heat, and shift flexible work into the cleaner hours of the grid. A Microsoft, Accenture, and WSP study cited by the Cloud for Global Good briefing puts the energy saving from moving on-premises software to the cloud at 30 to 90 percent. The top of the range is the best case, and where a given migration lands depends on its shape.

Uptime is sold as nines, paid back as credits

No provider sells constant uptime; the product is a monthly uptime percentage with service credits attached. The Amazon S3 service level agreement shows the standard shape: a credit worth 10 percent of the bill when monthly uptime lands under 99.9 percent, 25 percent under 99.0, and 100 percent under 95.0. The credit is the agreement’s “sole and exclusive remedy”: this refund is the whole remedy. S3’s own exclusions cover factors outside AWS’s reasonable control, from force majeure to internet trouble beyond the network’s demarcation point, the customer’s own actions and equipment, and account suspension or termination. Planned maintenance, the item cloud-SLA summaries usually list, does not appear in that agreement at all.

What do the nines add up to in downtime per year?
Commitment Approximate downtime per year
99 percent 3.7 days
99.9 percent 8.8 hours
99.99 percent 53 minutes
99.999 percent 5.3 minutes

Reading the nines as hours makes the ladder concrete: 99.9 percent tolerates roughly nine hours of downtime in a year, and 99.99 percent about fifty-three minutes. The arithmetic assumes a year of 8,760 hours. The agreement measures monthly windows rather than annual ones, so a single bad month can owe a credit inside an otherwise unremarkable year, and the percentages govern availability, a narrower thing than data durability, which S3 documents separately.


The invoice, item by item

Advertisements say no additional fees or hidden costs; the pricing page answers in a friendly font, and its component list runs storage, requests and data retrieval, data transfer, management and insights, replication, then transform and query. Requests are metered per thousand, a thousand GETs at $0.0004, and reads from the colder tiers, Standard-Infrequent-Access and Glacier, carry a retrieval charge on top of storage. Internet egress is free for the first 100 GB each month, an allowance aggregated across all AWS services and regions, China and GovCloud excepted, rather than an S3-only quota; beyond it S3 charges about $0.09 per GB in the Ireland example. S3 Intelligent-Tiering adds a monthly monitoring and automation charge per object kept there, and the management features meter by the million objects they cover. Storage itself bills the GB-month for whatever remains: the bucket a migration left behind accrues charges exactly like production data, until somebody empties it.

Which S3 line items surprise teams on a normal bill?
Line item What triggers it
Storage GB-months of data at rest; the row everyone expects
Egress Bytes leaving to the Internet; the first 100 GB a month is free account-wide (all AWS services combined, China and GovCloud excepted), then about $0.09 per GB in the Ireland example
Requests Billed per thousand; GET requests run $0.0004 per 1,000
Retrieval Reads from the Infrequent-Access and Glacier tiers, charged on top of storage
Monitoring and automation A monthly charge per object kept in the S3 Intelligent-Tiering class, for watching access patterns and moving objects between tiers
Idle allocations Resources nobody deleted; they bill until someone does

The table reads as a trap only when the shape of the workload is unknown. A static site with a CDN in front sees one row; a nightly full restore from cold storage meets several at once. The meter does not hide these lines; the brochure just declines to enumerate them.

Is encryption at rest automatic on S3?

The blanket claim that providers store your data encrypted flattens several distinct arrangements. Since January 5, 2023, Amazon S3 has applied default SSE-S3 encryption to every newly uploaded object, free of charge. The key-based alternatives, SSE-KMS and DSSE-KMS, wait for the customer to configure them. And responsibility for the data itself is shared across the provider line rather than transferred over it. Encryption is a badge earned by configuration, on both sides of that line.

What makes a cloud a cloud is the sentence NIST published in September 2011 plus the five characteristics attached to it, run as a checklist: the private data center stops at self-service and the meter, the dedicated server rented online in minutes at elasticity and the meter, and a service that meets all five earns the word regardless of who owns the hardware. The rest of the brochure, security and resilience included, scores how good a cloud is; the stamps alone settle membership.

Whiteboard board 15 of 15 titled Recap: a checked list reading a definition, not a logo; five stamps: self-service, network, pooling, elasticity, the meter; the brochure: outcomes; uptime: nines plus credits; the invoice outgrows the ad; encryption: configuration; beside a yellow sticky note reading In short, five stamps, pass or fail

Board 15 of 15: the recap checklist, with the sticky note carrying the operating rule.

Continue learning

The next lesson in this series, Cloud Trade-Offs, gives a full episode to what a pitch deck leaves unpriced, from the meter that never stops between traffic spikes to the cost of walking away.


Disclosure: this article was drafted with AI from the script of my video lesson, then checked against its sources before publishing.

This lesson first appeared on iqraa.tech.

More whiteboard lessons on the Whiteboard Architect YouTube channel.

Top comments (0)